GPS2

Christopher Janssen

Project Context:

Previous ARC studies required participants to manually provide detailed context about each GPS location they visited:

  • relation to location? (family/home/work)
  • protective/harmful to lapse?
  • access to alcohol?

This participant-provided information was then used to create predictive features for lapse probability.

Project Goal:

The GPS2 project investigates whether we can achieve similar predictive utility by extracting features from participants’ routine patterns and location metadata, all without requiring any manual contextualization.

This approach would ideally reduce participant burden while also potentially capturing behavioral patterns that participants themselves might not consciously recognize or report.

Data Pipeline and Workflow

The ideal workflow for GPS data is as follows:

  1. Classify Movement State
  2. Group Stationary Points into Clusters
  3. Enrich Clusters with Metadata
  4. Derive Predictive Features with Metadata

1. Movement Classification

  • Distinguish between
    • “stationary” points (where participants spend time)
    • “in movement” points (transitional GPS readings)
  • This allows us to focus on meaningful “activity stops” rather than travel paths

2. Location Clustering

  • Group nearby stationary GPS points into discrete location representations

    • Effectively combines any points within a building or radius into a single representative “place”

  • This allows us to create interpretable routine locations from raw coordinate clusters.

3. Metadata Enrichment

  • Apply geocoding services to automatically retrieve contextual information about each clustered location

    • business type

    • neighborhood characteristics

    • proximity to high-risk venues

    • etc

4. Feature Engineering

  • Transform location metadata and visit patterns into predictive features

    • time spent at different venue types

    • stability of work/housing

    • exposure to high-risk environments

  • All without requiring participant input!

The Granularity-Utility Paradigm

  • High granularity data > sophisticated behavioral analysis > significant privacy risks
  • Low granularity data > protects participant privacy > limits research capabilities
  • Inherent conflict between maximizing research utility and ensuring adequate privacy protection

Question:

What is the minimum level of location data granularity required to achieve valid research outcomes while maintaining acceptable privacy protection for participants?

Rundle et. al

Rundle et al.

  • Neighborhood health research increasingly uses GPS data + geocoding
  • Common practice is to send patient addresses to Google Maps, Census Bureau APIs
  • Each geocoding request = PII + PHI disclosure to third parties

GPS2 Infrastructure

  ┌─────────────────────────────────────────────────────────────────────┐
  │                         Local Computer                              │
  │                                                                     │
  │  ┌─────────────────┐    ┌──────────────────────────────────────┐    │
  │  │  Research Drive │    │           Docker Environment         │    │
  │  │                 │    │                                      │    │
  │  │  /Volumes/      │    │  ┌─────────────┐  ┌───────────────┐  │    │
  │  │  jjcurtin/      │────┼──│  PostGIS    │  │  Nominatim    │  │    │
  │  │  studydata/     │    │  │ Container   │  │  Container    │  │    │
  │  │  risk/          │    │  │             │  │               │  │    │
  │  │                 │    │  │ Port: 5433  │  │ Port: 8080    │  │    │
  │  │ • GPS data      │    │  │           <─│──┼>              │  │    │
  │  │ • Zoning Data   │    │  │ Database:   │  │ Service:      │  │    │
  │  │ • OSM Data      │    │  │ gps_analysis│  │ Reverse-      │  │    │
  │  └─────────────────┘    │  │             │  │ geocoding     │  │    │
  │                         │  │ User:       │  │               │  │    │
  │                         │  │ postgres    │  └───────────────┘  │    │
  │                         │  └─────────────┘                     │    │
  │                         └──────────────────────────────────────┘    │
  └─────────────────────────────────────────────────────────────────────┘

Local Processing Capabilities

  • All GPS data remains stored and processed on the researcher’s local machine throughout the entire analytical workflow
  • PostGIS enables sophisticated spatial operations without any external data transmission
  • Nominatim provides location metadata by converting coordinates to place names using locally hosted mapping data
  • Containerized architecture ensures reproducible analytical environments while maintaining complete data isolation from external networks

Current GPS2 Analytical Workflow

  ┌───────────────────────────────┐
  │        Quarto Notebooks       │
  │                               │
  │  01-setup-infrastructure.qmd  │
  │  02-data-import.qmd           │
  │  03-gps-processing-clustering │
  │  04-reverse-geocoding.qmd     │
  │  05-spatial-zoning-analysis   │
  │  06-visualizations.qmd        │
  └───────────────────────────────┘
                  │
                  ▼
  ┌─────────────────────────────────────────────────────────────┐
  │                       Integration                           │
  │  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────┐  │
  │  │      R      │◄─┤     SQL     ├─►│      PostGIS        │  │
  │  │             │  │             │  │                     │  │
  │  │ • dplyr     │  │ • Queries   │  │ • Spatial database  │  │
  │  │ • sf        │  │ • Joins     │  │ • Geographic data   │  │
  │  │ • leaflet   │  │ • Filtering │  │ • Spatial functions │  │
  │  │ • analysis  │  │ • Inserts   │  │ • Index operations  │  │
  │  └─────────────┘  └─────────────┘  └─────────────────────┘  │
  └─────────────────────────────────────────────────────────────┘

Connect to Database:

con <- connect_gps_db()

R-Database Integration + Visualization:

data <- pull_db(3, subid = 10)
data |> plot_cluster_map(color_by = "visits")

R-Database Integration + Visualization:

data <- pull_db(3, subid = 10)
data |> plot_cluster_map(color_by = "duration")

Zoning Visualization

pull_db(6) |> plot_zoning_map()

Database Disconnect

disconnect_gps_db(con)

Future Directions:

  1. Validate processing and clustering methods using statistical analyses
  2. Deep-dive into zone classification for the purposes of AUD
  3. Explore additional predictive metadata options from Nominatim